Multi-granularity contrastive learning model for next POI recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Next Point-of-Interest (POI) recommendation predicts user's next destination based on historical activities.
• Existing methods struggle with data sparsity using only location-level check-in trajectories.
• Leveraging region-level and category-level POI sequences can mitigate sparsity and improve recommendations.

Purpose of the Study:

• To propose a novel multi-granularity contrastive learning (MGCL) framework for next POI recommendation.
• To effectively utilize collaborative information across different granularities of POI sequences.
• To enhance user preference learning and recommendation performance by addressing data sparsity.

Main Methods:

• Constructing location-level POI graphs, category-level, and region-level sequences.
• Employing graph convolutional networks (GCNs) for cross-user sequential patterns and self-attention networks for individual user patterns.
• Applying contrastive learning to capture collaborative signals between multi-granularity sequences.
• Jointly training recommendation and contrastive learning tasks.

Main Results:

• The proposed MGCL method demonstrates superior performance compared to existing state-of-the-art methods.
• Multi-granularity representations effectively augment user preference learning.
• Contrastive learning successfully captures collaborative signals across different POI sequence granularities.

Conclusions:

• MGCL offers a significant advancement in next POI recommendation by integrating diverse data granularities.
• The framework effectively overcomes the limitations of data sparsity in traditional methods.
• Future work can explore further enhancements in capturing complex user behaviors and contextual information.